VisionStyle.

Detection results usually get drawn with whatever the library ships: a thin rectangle and a solid tag that look the same in a research figure, a security feed and a sports broadcast. Making them look intentional means rewriting OpenCV drawing code per project.

VisionStyle turns that into configuration. One Style covers boxes, strokes, fills, line patterns, labels, effects and trails, and the whole thing round-trips to YAML.

Twelve general-purpose looks and ten built for a specific job. Use one by name, or export its YAML and make it yours.

Sizes are in reference pixels at about 1080p and scale with the frame, so one style reads the same on a webcam and a 4K photo. On top of that, outlines and tags scale with each object's size, so distant objects stay light and close-up ones get heavier strokes.

visionstyle studio opens a local web app: pick a scene or upload your own image or video, run detection once, then tune every setting while scrubbing the timeline. The Python package renders every preview, so what you see is exactly what annotate() produces.

Saved looks become named presets you load anywhere with vs.Style.preset("my-look"), and the Export panel hands over the YAML, a Python snippet, or a prompt for AI coding tools.